Seen on X
What people are building with Laya.
Demos, games, ports and benchmarks posted on X, picked by hand. Each one links back to the original post and, where there is one, to the project in the directory.
LibLayaX 1.0.15 runs Laya models locally on CPU or GPU
LibLayaX 1.0.15 is out: an open-source alternative to the Jev API. It runs Laya decision models locally, from any language, on CPU or GPU. New: - a model in one .tar file - GPU under Wine, Apple Silicon included #OpenSource #AI #ML #LocalAI #Jev
Project: LibLayaX→
Laya Cybersec fine-tunes a model for German prompt-injection defense
A massive paradigm shift has happened in fine-tuning open-source models, I think @onusoz would agree. Training custom AI used to be insanely complex, but that barrier has been completely torn down. Frontier models are very capable but most of the time over kill. Most companies have repeating tasks that does not need GPT95-UltraProMax. We needed a specialized cybersecurity model to protect our users from prompt injection attacks. However, frontier models like @typesafeai Jev fell short on German language performance and couldn't be deployed locally in the EU. Even @Cloudflare CLEF is not able to match its performance. So, we built our own: • Trained on @runpod for under $100 • Leveraged 5 years of anonymized proprietary data • Achieved benchmark performance specialized frontier models can't touch • 100% data sovereignty with near-zero inference costs The future is open source. Dropping Laya Cybersec on @huggingface very soon! 🚀
Haystack adds TypeSafe document classification and text routing components
Most of what we ask LLMs to do inside pipelines isn't generation. It's decisions. @typesafeai System One models are built for exactly that, and they're now in @Haystack_AI The new integration gives you two components: 🏷️ TypeSafeDocumentClassifier answers your questions about each document and stores the results in its metadata. 🔀 TypeSafeTextRouter sends text to the right branch, with an optional low_confidence output so uncertain cases can fall back to an LLM or a human. Use them with TypeSafe's hosted Jev model, or keep everything on your own infrastructure with Ollaya, which serves open decision models like Laya locally through the TypeSafe API. 🐍 pip install typesafe-haystack 👉
Hedos 1.5.0 adds a redesigned TUI for running Laya and other decision models
released hedos 1.5.0 today. now you can pull and run decision models like cloudflare's clef, openjev, kev, laya and the rest. plus it comes with the new redesigned tui.
laya-mlx delivers typed decisions in 7–14 ms on an M3 Max
Vous vous êtes déjà retrouvé confronté à des décisions rapides qui nécessitent une inférence rapide et efficace ? Découvrez Laya-mlx, un environnement d'exécution MLX natif pour les modèles de décision typés Laya – des décisions courtes de 7 à 14 ms sur M3 Max. Sans génération de texte ni PyTorch ou API cloud, il offre des décisions typées à poids ouvert exécutées nativement sur Apple Silicon.
Project: laya-mlx→
Free Unsloth notebook trains Laya-style decision models with LoRA
🔥颠覆｜普通LLM变身决策AI，免费Notebook可用 免费训练决策模型，小显存就能跑！ Unsloth推出Colab/Kaggle免费Notebook，可把Qwen、Gemma这类大模型改成Jev/Laya/Clef风格决策模型，不再生成长篇文本，直接输出选项+置信概率。 仅用LoRA微调1个epoch，基准准确率从30%飙升至最高78%。4GB显存就能跑0.8B版本，8GB显存即可训练Qwen3.5-4B。 适配choice、noul、score三类任务，训练后支持本地部署，数据不外流。 工单分配、风控、内容审核场景直接可用，上手门槛极低。 你打算拿这种本地决策模型落地什么场景？
Laya scores 44% accuracy on the MASSIVE intent set
Tried Liquid's new d1-3B locally (Q4_K_M) against Laya on the full MASSIVE intent set. 1. d1-3B 85% accuracy, 14 ms per decision. When wrong, confidence drops to 62%. 2. Laya 44% accuracy. When wrong, still 90% sure. I'll take the model that admits it's guessing. 👀
STAGE runs Laya in production for decision tasks
We've been running Laya in production at STAGE for all sorts of decision work. Already on to trying Vega pipelines. Indian builders shipping at this level is the real story. So so proud - we need to hype this much more!
A Jev-style model fine-tuned from Laya with Firecrawl data
I built my own version of Jev in 3 hours b fine tuning laya with Firecrawl data. The results are pretty impressive...
A local Laya proof of concept classifies support tickets
Fiz uma POC no trabalho de usar laya (parecido com jev só que opensource e local) pra classificar tickets de suporte. Vamo ver se vai ser útil.
A livestream tests how Laya handles an infinite-trolley prompt
Riding Infinite Trolly - Tricking Decision models + Jev, Clef, Jev, Kev and Laya etc
sys1rust runs Laya decision models up to 4.1× faster on an M5
most system 1 models are doing math its output never used i built sys1rust to run laya's decision models locally in rust. up to 4.1x faster than stock laya serve on my base m5 macbook pro the video takes you inside the gpu to show how text becomes a decision, and where i found the speedup code below!
Laya beats Jev at Tetris in a local gameplay demo
Chinese Laya casually beating JEV on Tetris.
Jev and Laya are compared against Opus on Rubik’s Cube moves
@aimlapi Opus taking 72 seconds for just three moves shows it actually built a spatial model of the cube. Jev and Laya just panicked and spammed API calls until they hit the limit.
Chinese Laya fine-tune raises target-class agreement from 34% to 76%
把 Laya multilingual 决策模型的中文能力微调出来：翻译公开英文决策数据、保留软目标分布，训练后目标类别一致率从 34% 提到 76%。
Project: chinese-laya→
Ollaya benchmarks local Laya decisions against hosted Jev
Ollaya runs decision models locally: typed, calibrated answers in one forward pass, no token generation. Five questions to laya in 8–10 ms on an RTX 4090 vs 236–276 ms median for TypeSafe's hosted Jev API. TypeSafe SDK 0.7.1 runs unchanged against localhost:11435.
A local Laya model validates decisions in a coding pipeline
Testing Laya (a tiny local decision model) in my coding pipeline as a decision validator. Sometimes it contradicts itself, and the 'confidence' is the way to filter if it is possible to use its answer or not.
Ollaya offers a local API for Laya and other decision models
Ollaya just killed TypeSafe Jev — a free, self-hosted, open source alternative that runs locally on your own machine, no API key required. @typesafeai 's Jev kicked off a whole new category of AI model built for fast, typed decisions instead of generated text — and weeks later, Ollaya cloned it into something you can self host for free. This breakdown covers what a decision model actually is, why regular LLMs are a slow and expensive way to answer a simple yes or no question, how Jev's Choice, Score, and Noul questions work under the hood, and exactly how Ollaya recreates that same TypeSafe-compatible API locally so existing code just works against your own machine. You'll also get a full walkthrough of every model Ollaya supports — Laya, Decider, NLI, and Gliclass — and which one actually makes sense depending on whether you care more about speed or accuracy. If you're a developer looking for a TypeSafe Jev alternative that's open source, privacy-first, and doesn't charge per request, this one's for you. Whether you're building support ticket routing, content moderation, sentiment scoring, or any kind of local AI decision-making pipeline, this video breaks down whether Ollaya actually holds up against the real thing — including an honest look at its accuracy claims, its calibration numbers, and where the project still shows its age as a brand new open source release. @GithubProjects @github @mertcobanov
Laya LiteRT sorts 120 unread texts on a Galaxy S26
laya converted to LiteRT. A decision encoder (ModernBERT-large, Apache-2.0): a text and options in your words in, an answer with its probability out. Clip: 120 unread texts on a Galaxy S26 sorted by what each needs from you, 133 ms a text, 16.6 s in all.
Project: laya-LiteRT→
Fine-tuning Laya for skill routing
Laya: Fine-tuning the Open Jev-style Decision Model on Skill Routing via @YouTube
Video explains Jev and Laya for fast, low-cost decisions
Yapay zekada her karar için para ödemek şart mı? Jev, küçük kararları hızlı ve ucuza vermeyi hedefliyor. Laya ise benzer işleri kendi bilgisayarınızda yapabilmeniz için açık modeller sunuyor. Peki bu bizim için ne değiştirir? Kısaca anlattım. 👇
Explaining how Laya can classify mobile app reviews
Typesafe'in JEV modeli nedir? FİKİR: mobil oyun uygulama geliştiriyorsanız, kullanıcılarınız varsa, JEV (veya açık kaynak alternatifi LAYA'yı) kullanarak şunu yapabilirsiniz: en basitinden, uygulamanızın google play yorumlarının tamamını çekip bunları kaydedersiniz. ardından seçenekler belirlersiniz (JEV 255 seçeneğe kadar destekliyor) - bu kullanıcı kızgın mı? - bu kullanıcı memnun mu? ... seçenekler sizin ihtiyacınıza göre çoğaltılabilir. bunların hepsine karar vermede JEV hem çok hızlıdır hem çok isabetli. size basitçe evet/hayır cevabı döner. uzun uzadıya yazmaz, zaten bilmez. JEV'in ne olduğunu hem bir kullanım senaryosu üzerinden anlatmak, hem de bir fikir vermek amaçlı yazıyorum. kullandığınız modeller gün içerisinde bir sürü karar veriyor ve çoğunda bu iş hem uzun sürüyor hem de pahalı (kullanım kotanızı yiyor). JEV veya alternatifi modeller bu işi hem çok ucuza (açık kaynak LAYA 2-3 GB boyutu var, klasik bi bilgisayarda bile çalışır) hem de daha tutarlı şekilde yapabilir
LiteLLM adds support for Laya and Nimble classifiers
Thrilled to support OPEN-WEIGHT DECISION MODELS on @LiteLLM Laya + Nimble are now live on LiteLLM
Nine Android demos run Laya on Acurast, from games to prompt-injection defense
Nine live demos put Laya, an open System One model, on a real Android phone in the Acurast Network. Snake, Tetris, Doom, mail sorting, and a prompt-injection guard. Try it at Walkthrough from @Acurast. 📱
How classification evolved from word counts to decision models
Jev, Laya, Span-01, ScaleDown Classify...everyone's talking about decision models But classification has been around forever We broke down the 3 types (binary, multiclass, multilabel) and how the tech evolved from word counts > BERT > LLMs > decision models
laya-mlx plays Snake at 60 decisions per second on an M3 Max
Introducing a version 50 times faster than Jev, running on your device: laya-mlx! Occupies a maximum of 1G memory only on your device Laya is an open-source classification system similar to Jev, based on text output probabilities I ported it to MLX and made some performance optimizations! In the video, this model is playing Snake on my local M3 Max This model can play Snake at a decision speed of 60 times per second!
Pi adds llama.cpp support for local Laya decision classification
🔥 Pi v1.1.0 更新速递：增量工具微调、终端状态感知与决策分类模型 这次带来了工具集增量微调、OSC 终端状态汇报以及多款决策分类模型支持，并重点改善了 MCP 交互与调试输出体验。 1. 工具与终端状态 · --tools 支持通过 +name 与 -name 增量微调默认工具，不再全量覆盖 例：pi -t +codemode,-write 在默认工具集基础上增删 · 支持 OSC 7501 状态协议，终端和面板可实时感知 Pi 处于运行、等待交互、已完成还是失败 · agent_settled 事件新增 aborted 状态标记，便于外部集成区分任务取消与正常结束 · pi update 自动清理冗余历史版本，仅保留当前版本和上一版本 2. 模型与决策分类 · 新增 anthropic/claude-haiku-5-5，支持最高 xhigh/max 档位自适应思考及 Bedrock 提示词缓存 · Decisions API 接入 GPT-6 Luna 决策分类模型，codemode 的 models.classify() 支持传入图片判断 · 原生接入 llama.cpp 0.6.0+ 本地决策模型（Julia-1、Laya 等），通过 /v1/systemone 作为纯分类器调用 · 遇到服务商返回 server_busy 繁忙错误时自动重试，不再直接中断当前轮次 3. MCP 与交互细节 · /mcp 管理器改为实时更新，不再卡死等待全部服务端连通才打开 · MCP OAuth 登录支持随时按 Esc 取消，请求增加 15 秒超时，避免退出时后台残留挂起 · codemode 多段输出采用 ==> text N/M <== 显式分块，console 输出统一聚合，避免模型混淆内容 · 修复 Anthropic 浏览器登录在 53692 端口被占用时报错的问题，自动回退到可用随机端口
BI conversational-agent case study compares Laya, Jev, and LLM decisions
O LLM é normalmente o cérebro de um agente analítico: interpreta a pergunta, escolhe a ferramenta, gera a consulta e explica o resultado. O Estudo de Caso 1 de BI Conversacional e Construção de Agentes Analíticos olha para um detalhe que costuma passar despercebido nessa descrição. Antes de qualquer consulta, o agente toma uma série de decisões pequenas, repetitivas e com um conjunto fechado de respostas: a pergunta é sobre um indicador ou sobre um documento? Qual indicador? Falta alguma informação? A solicitação traz algum risco? Usar um modelo de linguagem completo para cada uma delas custa tempo, dinheiro e exige interpretar o texto devolvido. Uma categoria de modelos surgiu exatamente para esse tipo de decisão: os System One models, dos quais o Jev (TypeSafe AI) e o Laya (Convai Innovations) são os principais representantes. Em vez de gerar texto, eles recebem um estado e perguntas tipadas e devolvem decisões com probabilidades calibradas, em milissegundos. O Estudo de Caso compara esses dois modelos com o LLM na tomada de decisões analíticas de um copiloto de BI conversacional. Bons estudos.
Posted a Laya demo on X? Mention “Laya” and it gets picked up automatically, or submit the project.